Teaching robots through trial and error

A few days after watching a small army of Sawyer robots drop fake apples into plastic bowls, we’re back at UC Berkeley’s Sutardja Dai Hall to witness another approach to robotic learning. This time out, it’s an older model — PR2, the once ubiquitous personal robot from the now-defunct Willow Garage.

The UCB research team has deemed the robot BRETT — that’s the Berkeley Robot for Elimination of Tedious Tasks. Like the Sawyer robots, BRETT is here to learn — and hopefully offer researchers some valuable insight into how we can teach robots how to perform dull and repetitive tasks without a lot of programming.

In this case, the job is picking and placing — a decidedly tedious warehouse task that has become extremely demanding as online retailers like Amazon have put the crunch on logistics companies. In some scenarios, robots are taught to execute the process through human demonstrations. In others, they’re programmed with tens of thousands of computer simulations.

BRETT, on the the other hand, is programmed to learn by doing — and correcting its own mistakes. “We’re putting robots in situations where there’s so much variation that it’s impractical to program ahead of time every possible situation for the robot,” explains Pieter Abbeel, the Director of Berkeley’s Robot Learning Lab. “You want it to learn from its own experience.”

In a demo, BRETT is asked to perform a preschool puzzle, placing a block through a hole in wooden box. The robot’s first several attempts are clumsy, missing the box entirely. Through trial and error, however, the robot gets closer, finally accomplishing the task.

“It’s like the way a dog is trained,” says Abbeel. “You say ‘good dog’ or ‘bad dog.’ That’s what’s essentially what’s happening. From that feedback, it starts fine tuning its skill and getting that block into a matching opening.”

Pieter Abbeel will be appearing at TC Sessions: Robotics next week at U.C. Berkeley.